1. - 1 -

2. - 2 - Case studies in recommender systems  The MovieLens data set, others –focus on improving the Mean Absolute Error …  What about the business value? –nearly no real-world studies –exceptions, e.g., Dias et al., 2008.  e-Grocer application  CF method  short term: below one percent  long-term, indirect effects important  This study –measuring impact of different RS algorithms in Mobile Internet scenario –more than 3% more sales through personalized item ordering

3. - 3 - Application platform  Game download platform of telco provider –access via mobile phone –direct download, charged to monthly statement –low cost items (0.99 cent to few Euro)  Extension to existing platform –"My recommendations" –in-category personalization (where applicable) –start-page items, post-sales items  Control group –natural or editorial item ranking –no "My Recommendations"

4. - 4 - Study setup  6 recommendation algorithms, 1 control group –CF (item-item, SlopeOne), Content-based filtering, Switching CF/Content- baaed hybrid, top rating, top selling  Test period: –4 weeks evaluation period –about 150,000 users assigned randomly to different groups –only experienced users  Hypothesis (H1 – H4) –H1: Pers. recommendations stimulate more users to view items –H2: Person. recommendations turn more visitors into buyers –H3: Pers. recommendations stimulate individual users to view more items –H3: Pers. recommendations stimulate individual users to buy more items

5. - 5 - Measurements  Click and purchase behavior of customers –customers are always logged in –all navigation activities stored in system  Measurements taken in different situations –my Recommendations, start page, post sales, in categories, overall effects –metrics  item viewers/platform visitors  item purchasers/platform visitors  item views per visitor  purchases per visitor  Implicit and explicit feedback –item view, item purchase, explicit ratings

6. - 6 - My Recommendations conversion rates Item viewers / visitorsPurchasers / visitors  Conversion rates –top-rated items (SlopeOne, Top-Rating) appear to be non-interesting –only CF-Item able to turn more visitors into buyers (p < 0.01)  Overall on the platform –no significant increase on both conversion rates (for frequent users!)

7. - 7 - My Recommendations sales increase (1) Item views/ customerPurchases / customer  Item views: –except SlopeOne, all personalized RS outperform non-personalized techniques  Item purchases –RS measurably stimulate users to buy/download more items –content-based method does not work well here

8. - 8 - My Recommendations sales increase (2)  Demos and non-free games: –previous figures counted all downloads –figure shows  personalized techniques comparable to top seller list  however, can stimulate interest in demo games  Note, Rating possible only after download Figure shows purchases per visitor rate Note: Only 2 demos in top 30 downloads

9. - 9 - Post-sales recommendations Item views / visitorPurchases / visitor  Findings –recommending "more-of-the-same", top sellers or simply new items does not work well –top-Rating and SlopeOne nearly exclusively stimulate demo downloads (Not shown) –top-Seller und control group sell no demos

10. - 10 - Start page recommendations  Example: –purchasers / visitors conversion rate  Findings: –visual presentation is important, click distribution as expected (omitted here) –personalization raises attraction also on text links –proposing new items works also very well on the start pages

11. - 11 - Overall effects  Overall number of downloads (free + non-free games)  Pay games only Notes In-category measurements not shown in paper. Content-based method outperforms others in different categories (half price, new games, erotic games) Effect: 3.2 to 3.6% sales increase!

12. - 12 - Further observations: ratings on the Mobile Internet  Only 2% of users issued at least one rating –most probably caused by size of displays –in addition: Particularity of platform; rating only after download –insufficient coverage for standard CF methods  Implicit ratings –also count item views and item purchases –increase the coverage of CF algorithms –MAE however not a suitable measure anymore for comparing algorithms

13. - 13 - Summary  Large case study on business effects of RS –significant sales increase can be reached! (max. 1% in past with other activities) –more studies needed –value of MAE measure …  In addition –recommendation in navigational context  acceptance of recommendation depends on situation of user  Further work –comparison of general sales behavior –more information in data to be found